SmallInitEmb - LayerNorm(SmallInit(Embedding)) in a Transformer to improve convergence

Overview

SmallInitEmb

LayerNorm(SmallInit(Embedding)) in a Transformer

I find that when training a transformer, the embedding matrix moves slowly, hence it's difficult for the model to jump out of the initial noisy embedding.

(initial embedding)
[[-0.0073  0.0062 -0.0261 ...  0.0086  0.0107 -0.008 ] ... ]
 (after 1 step, the directions of the embedding vectors are not moved much because the numbers change by ~LR = ~4e-4)
[[-0.0069  0.0066 -0.0265 ...  0.009   0.0111 -0.0084] ... ]

So I propose initializing the embedding matrix to tiny values, and put another LayerNorm after it (before all the SA & FFN layers):

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
if self.config.USE_SMALL_EMB and self.layer_id == 0:
    x = self.lnPre(x) # LN(SmallInit(Emb))
x = x + self.att(self.ln1(x))
x = x + self.ffn(self.ln2(x))

And then you get improved convergence (especially for BPE models) because the model can quickly jump out of the tiny initial embedding (small changes after 1 step -> significant changes of directions -> significant changes after LayerNorm).

Loss curve comparison: https://wandb.ai/blinkdl/SmallEmbTest

(the gap between LayerNorm(SmallEmb)) and baseline persists after more training)

Moreover, you can directly train PostLN models without warmup with SmallInit(Emb)

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
x = self.ln1(x) # this plays the same role as the lnPre in the above PreLN code
x = x + self.att(x)
x = self.ln2(x)
x = x + self.ffn(x)
(note you shall have another LN after the final ffn)
Owner
PENG Bo
http://zhihu.com/people/bopengbopeng
PENG Bo
Semantic Bottleneck Scene Generation

SB-GAN Semantic Bottleneck Scene Generation Coupling the high-fidelity generation capabilities of label-conditional image synthesis methods with the f

Samaneh Azadi 41 Nov 28, 2022
A fast poisson image editing implementation that can utilize multi-core CPU or GPU to handle a high-resolution image input.

Poisson Image Editing - A Parallel Implementation Jiayi Weng (jiayiwen), Zixu Chen (zixuc) Poisson Image Editing is a technique that can fuse two imag

Jiayi Weng 110 Dec 27, 2022
TensorFlow 101: Introduction to Deep Learning for Python Within TensorFlow

TensorFlow 101: Introduction to Deep Learning I have worked all my life in Machine Learning, and I've never seen one algorithm knock over its benchmar

Sefik Ilkin Serengil 896 Jan 04, 2023
RNG-KBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering

RNG-KBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering Authors: Xi Ye, Semih Yavuz, Kazuma Hashimoto, Yingbo Zhou and

Salesforce 72 Dec 05, 2022
Lingvo is a framework for building neural networks in Tensorflow, particularly sequence models.

Lingvo is a framework for building neural networks in Tensorflow, particularly sequence models.

2.7k Jan 05, 2023
Unsupervised Discovery of Object Radiance Fields

Unsupervised Discovery of Object Radiance Fields by Hong-Xing Yu, Leonidas J. Guibas and Jiajun Wu from Stanford University. arXiv link: https://arxiv

Hong-Xing Yu 148 Nov 30, 2022
Code and Data for the paper: Molecular Contrastive Learning with Chemical Element Knowledge Graph [AAAI 2022]

Knowledge-enhanced Contrastive Learning (KCL) Molecular Contrastive Learning with Chemical Element Knowledge Graph [ AAAI 2022 ]. We construct a Chemi

Fangyin 58 Dec 26, 2022
HairCLIP: Design Your Hair by Text and Reference Image

Overview This repository hosts the official PyTorch implementation of the paper: "HairCLIP: Design Your Hair by Text and Reference Image". Our single

322 Jan 06, 2023
Code for NeurIPS 2021 paper: Invariant Causal Imitation Learning for Generalizable Policies

Invariant Causal Imitation Learning for Generalizable Policies Ioana Bica, Daniel Jarrett, Mihaela van der Schaar Neural Information Processing System

Ioana Bica 17 Dec 01, 2022
TSIT: A Simple and Versatile Framework for Image-to-Image Translation

TSIT: A Simple and Versatile Framework for Image-to-Image Translation This repository provides the official PyTorch implementation for the following p

Liming Jiang 255 Nov 23, 2022
Tiny Kinetics-400 for test

Kinetics-400迷你数据集 English | 简体中文 该数据集旨在解决的问题:参照Kinetics-400数据格式,训练基于自己数据的视频理解模型。 数据集介绍 Kinetics-400是视频领域benchmark常用数据集,详细介绍可以参考其官方网站Kinetics。整个数据集包含40

38 Jan 06, 2023
Unsupervised Image-to-Image Translation

UNIT: UNsupervised Image-to-image Translation Networks Imaginaire Repository We have a reimplementation of the UNIT method that is more performant. It

Ming-Yu Liu 劉洺堉 1.9k Dec 26, 2022
Deep learning with dynamic computation graphs in TensorFlow

TensorFlow Fold TensorFlow Fold is a library for creating TensorFlow models that consume structured data, where the structure of the computation graph

1.8k Dec 28, 2022
Cave Generation using metaballs in Blender. Originally created by sdfgeoff, Edited by Myself (Archie Jaskowicz).

Blender-Cave-Generation Cave Generation using metaballs in Blender. Originally created by sdfgeoff, Edited by Myself (Archie Jaskowicz). Installation

2 Dec 28, 2022
Code for the paper "Improving Vision-and-Language Navigation with Image-Text Pairs from the Web" (ECCV 2020)

Improving Vision-and-Language Navigation with Image-Text Pairs from the Web Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh

Arjun Majumdar 44 Dec 14, 2022
State of the Art Neural Networks for Deep Learning

pyradox This python library helps you with implementing various state of the art neural networks in a totally customizable fashion using Tensorflow 2

Ritvik Rastogi 60 May 29, 2022
Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces(ICML 2021)

Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces(ICML 2021) This repository contains the code

149 Dec 15, 2022
[ICCV 2021] Code release for "Sub-bit Neural Networks: Learning to Compress and Accelerate Binary Neural Networks"

Sub-bit Neural Networks: Learning to Compress and Accelerate Binary Neural Networks By Yikai Wang, Yi Yang, Fuchun Sun, Anbang Yao. This is the pytorc

Yikai Wang 26 Nov 20, 2022
基于YoloX目标检测+DeepSort算法实现多目标追踪Baseline

项目简介: 使用YOLOX+Deepsort实现车辆行人追踪和计数,代码封装成一个Detector类,更容易嵌入到自己的项目中。 代码地址(欢迎star): https://github.com/Sharpiless/yolox-deepsort/ 最终效果: 运行demo: python demo

114 Dec 30, 2022
Real-world Anomaly Detection in Surveillance Videos- pytorch Re-implementation

Real world Anomaly Detection in Surveillance Videos : Pytorch RE-Implementation This repository is a re-implementation of "Real-world Anomaly Detectio

seominseok 62 Dec 08, 2022